OpenAI CEO Sam Altman is lobbying the Biden administration to support the development of a vast network of AI datacenters, each with a power requirement of up to five gigawatts. This initiative is deemed crucial for US national security and to maintain technological supremacy over China.

A document reviewed by Bloomberg outlines the plan to establish multiple datacenters across the US. The push comes just weeks after Altman and other technology leaders convened at the White House to discuss the future of AI infrastructure.

Constellation Energy Corp. executives hinted that the initial phase might see the construction of 5-7 datacenters, starting with one. Building such a facility, however, poses significant challenges. The power needed is comparable to that produced by five pressurized water nuclear reactors.

The proposed datacenters would be second only to the Grand Coulee Dam in Washington in terms of power generation. Other large plants like Georgia’s Alvin W. Vogtle and Arizona’s Palo Verde would fall below these new facilities in output.

Already, power capacity constraints are delaying many datacenter projects. A CBRE report highlighted a shortage of both power and the equipment to manage it, causing project delays.

Cloud providers are increasingly using unconventional energy sources to stay competitive in AI development. Microsoft recently entered a 20-year agreement with Constellation Energy to resurrect the Three Mile Island Unit 1 nuclear plant, while Amazon has secured up to 960 megawatts of power through a $650 million deal with Talen Energy’s Susquehanna plant.

Oracle has even considered installing small modular reactors to power its datacenters, reflecting a broader trend towards innovative energy solutions.

Meeting the power demands is one hurdle; another is acquiring enough hardware. A five-gigawatt facility could potentially house over 35,000 Nvidia Grace-Blackwell rack-scale systems, equating to roughly 2.5 million Blackwell GPUs.

Nvidia shipped approximately 600,000 H100 chips last year and aims to ship up to 2 million this year, but TSMC’s limited advanced packaging capacity adds complexity to scaling production.

Altman is known for ambitious proposals, such as a previously suggested $7 trillion network of chip factories. Thus, the AI datacenter plan may be more about stimulating government investment in AI than immediate action.